CS-Pi: Constraint-Satisfying Flow Policy Sampling
CS-Pi: Constraint-Satisfying Flow Policy Sampling
This paper introduces CS-Pi, a novel sampling technique for flow-matching robot policies that enforces safety constraints at inference time without retraining. The method projects actions to the nearest safe state and adds noise scaled by the projection distance before resampling, balancing efficiency with action plausibility. Experiments on a 7-DoF shelf pick-and-place task and zero-shot Franka hardware deployment show CS-Pi achieves 81% and 55% safe success rates respectively, outperforming standard projection (55% and 40%) while maintaining plausible action distributions.
The authors propose CS-Pi, a sampling method that combines projection and rejection to enforce safety constraints at inference time while maintaining action plausibility for flow-matching policies, demonstrating improved safe success rates on both simulation and hardware benchmarks compared to baseline projection methods.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
DoF of shelf pick-and-place
7
View original evidence
ejection's plausibility. CS-Pi, our draw-project-renoise sampler, projects each draw to the nearest safe action, adds noise back scaled to how far the projection moved it, and re-runs the unmodified policy until a draw is safe. Replacing only the projection step extends CS-Pi to continuous costs. Across a 2D diagnostic, a 7-DoF shelf pick-and-place, an orientation preference, and a zero-shot Franka deployment, CS-Pi
Open source S5diagnostic dimensionality
2
View original evidence
ejection's plausibility. CS-Pi, our draw-project-renoise sampler, projects each draw to the nearest safe action, adds noise back scaled to how far the projection moved it, and re-runs the unmodified policy until a draw is safe. Replacing only the projection step extends CS-Pi to continuous costs. Across a 2D diagnostic, a 7-DoF shelf pick-and-place, an orientation preference, and a zero-shot Franka deployment, CS-Pi
Open source S5
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ejection's plausibility. CS-Pi, our draw-project-renoise sampler, projects each draw to the nearest safe action, adds noise back scaled to how far the projection moved it, and re-runs the unmodified policy until a draw is safe. Replacing only the projection step extends CS-Pi to continuous costs. Across a 2D diagnostic, a 7-DoF shelf pick-and-place, an orientation preference, and a zero-shot Franka deployment, CS-Pi
Open source S5
reaches 81% safe success on the shelf and 55% on hardware, against 55% and 40% for projection, while staying plausible.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org